PcPreT-Net: Predicting classification of decline rate in prostate-specific antigen using graph neural network

Abstract

Prostate cancer (PCa) is one of the most common cause of cancer-related deaths among men worldwide, with prostate-specific antigen (PSA) serving as a widely accepted biomarker for the diagnosis, treatment monitoring, and prognosis of PCa. Accurate assessment of PSA dynamics is therefore essential for evaluating therapeutic efficacy and disease progression. Magnetic resonance imaging (MRI) is widely recognized for its accuracy and non-invasive nature in managing PCa, plays a key role in PCa management. We aim to establish a predictive association between MRI data and PSA decline to enable individualized treatment assessment. This study proposes a hybrid classification model combing convolutional neural network (CNN) and graph convolutional network (GCN) to predict PSA decline rate. The graph nodes are constructed from multiparametric MRI (mp-MRI) images with highlighting tumor regions. The CNN, pretrained to classify Gleason score risk levels, serves as an image feature extractor that extracts semantic features and encodes inter-node relationships. Based on these features, a mapping relationship between mp-MRI and PSA decline rate categories was then developed. Ablation experiments validated the effectiveness of the designed feature extraction framework. Comparative tests showed that our model outperformed traditional radiomics, CNN, and vision transformer (ViT) models, achieving an accuracy of 0.870, precision of 0.881, recall of 0.858, and F1-score of 0.872.